TorchCRF 1.0.4

An Implementation of Conditional Random Fields in pytorch

Torch CRF

Implementation of CRF (Conditional Random Fields) in PyTorch 1.0

Requirements

• python3 (>=3.6)
• PyTorch 1.0

Installation

$ pip install TorchCRF

Usage

>>> import torch
>>> from TorchCRF import CRF
>>> batch_size = 2
>>> sequence_size = 3
>>> num_labels = 5
>>> mask = torch.FloatTensor([[1, 1, 1], [1, 1, 0]]) # (batch_size. sequence_size)
>>> labels = torch.LongTensor([[0, 2, 3], [1, 4, 1]])  # (batch_size, sequence_size)
>>> hidden = torch.randn((batch_size, sequence_size, num_labels), requires_grad=True)
>>> crf = CRF(num_labels)

Computing log-likelihood (used where forward)

>>> crf.forward(hidden, labels, mask)
tensor([-7.6204, -3.6124], grad_fn=<ThSubBackward>)

Decoding (predict labels of sequences)

>>> crf.viterbi_decode(hidden, mask)
[[0, 2, 2], [4, 0]]

License

MIT

References

• threelittlemonkeys/lstm-crf-pytorch
• kmkurn/pytorch-crf